The short answer
Hotword biasing is a way to give a speech recognition model a short list of names and terms to expect, so that it favours them when the audio sounds close, without retraining the model. It fixes the most visible class of errors: product names, people, places and jargon that a general model spells wrong.
Good biasing only tips close calls. A listed word that was not said should stay out of the transcript, and without a list the output should not change.
Ways to teach a model new words
Biasing at decode time is one of several approaches. They differ in when they act and what they cost.
| Method | When it acts | Retraining | Handles a word the model never saw | Main risk |
|---|---|---|---|---|
| Hotword list at decode time | During decoding, per request | No | Often, if the spelling can be built from the model's tokens | Too strong a bonus inserts listed words that were not said |
| Contextual language model or class grammar | During decoding, from an external language model | The language model only | Yes, if added to the grammar | Extra component to maintain |
| Neural contextual biasing (for example CLAS) | Inside the model, from embeddings of the phrase list | Yes, the model is trained to use lists | Yes | Needs a model built for it |
| Fine-tuning on domain audio | Before deployment | Yes | Yes, if the data contains it | Cost, and drift on general speech |
How decode-time biasing works
A decoder keeps several candidate transcripts and scores each one as it goes. With a hotword list, a candidate whose latest tokens spell the start of a listed term gets a small score bonus, and keeps it if the term is completed. A weight controls how big the bonus is: too small and nothing changes, too large and the model starts hearing listed words in unrelated audio.
Neural contextual biasing goes further. Pundak et al. ("Deep context", arXiv 1808.02480, 2018) trained a Listen, Attend and Spell model jointly with embeddings of the context phrases, and reported up to 68 % relative WER reduction over a baseline that combined separate models by shallow fusion.
Example: hotwords in Phonon-2
Phonon-2 takes up to 25 hotwords on every engine it runs on: Apple silicon, the CPU engine on Linux, Windows and macOS, and the CUDA image. Multi-word terms such as Ada Lovelace are kept whole. --hotword-lambda sets how hard listed words pull, with a default of 2.0. A listed word that was not said stays out of the transcript.
On the server, send a hotwords field per request, or the OpenAI prompt field, which is read as a vocabulary list. Hotwords apply only to the request that sends them. Hotwords are not part of the current Phonon-2.5 release; today they are a Phonon-2 feature.
Practical tips: list the exact spellings you want to see, keep the list to terms that are likely in this audio, and measure a sample of your own recordings with and without the list before turning the weight up.
fermion transcribe phonon-2 call.wav --hotwords "Ada, Quillon"
fermion transcribe phonon-2 call.wav --hotwords @names.txt
curl -s http://127.0.0.1:8000/v1/audio/transcriptions -F file=@call.wav -F model=phonon-2 -F "hotwords=Ada, Quillon"Questions
- What is hotword biasing in speech recognition?
- Giving the model a list of expected names and terms at transcription time so it favours them when the audio is close. It needs no retraining.
- Is hotword biasing the same as custom vocabulary?
- Mostly. Custom vocabulary, keyword boosting and contextual biasing are names different vendors use for the same goal; the method underneath can be a decode-time bonus, an external language model or a model trained to read phrase lists.
- Can hotwords make a model hear words that were not said?
- Yes, if the bonus is too strong. A well-tuned list only tips close calls; in Phonon-2 a listed word that was not said stays out of the transcript.
- How many hotwords can I give Phonon-2?
- Up to 25 terms per request, comma-separated or from a file with @names.txt. Extra terms sent to the server are dropped with a note.
- Does Phonon-2.5 support hotwords?
- Not in the current release. Hotwords are available on Phonon-2 today, from the command line, Python and the server.
- Does the OpenAI prompt field work as a hotword list?
- On the Phonon server, yes: the prompt field is read as a vocabulary list when no hotwords field is sent.
Related
Sources and method
- Pundak, Sainath, Prabhavalkar, Kannan, Zhao, Deep context: end-to-end contextual speech recognition, arXiv 1808.02480 (2018-08-07), read 2026-11-13.
- Phonon-2 hotwords: Fermion Research docs (/docs/speech/, /docs/cli-reference/, /docs/server-reference/) and github.com/fermionresearch/phonon docs/hotwords.md, read 2026-11-13.
Canonical: https://www.fermionresearch.com/glossary/hotword-biasing/